Warm-Up Period

The primary function of a warm-up period is to populate the historical lookback windows required by technical indicators and state variables. A 200-day moving average, for example, requires 200 data points before it generates a valid value. Without a warm-up period, the initial signals are calculated on incomplete data, producing NaN values or extreme outliers that corrupt the equity curve. The engine processes all events during this phase but suppresses performance logging.

For strategies with path dependency, the warm-up period establishes the initial state of the decision-making automaton. This includes:

• Populating rolling window caches for volatility estimators
• Seeding the initial capital allocation and position vectors
• Calibrating dynamic thresholds for signal generators
• Warming up online learning models that adapt to streaming data The state at the end of the warm-up window becomes the deterministic starting condition for the evaluation phase.

The warm-up period creates a strict temporal boundary in the dataset. Data before the boundary is the burn-in segment, used exclusively for populating indicators. Data after the boundary is the evaluation segment, where all performance metrics are recorded. This split must be enforced at the engine level to prevent information leakage—a single bar of overlap can contaminate the Sharpe Ratio calculation by including a trade that used incomplete indicator values.

The required warm-up duration is not fixed; it is a function of the strategy's maximum lookback window. A mean-reversion strategy using a 50-day Bollinger Band needs 50 bars. A multi-timeframe strategy combining daily and weekly signals may need the least common multiple of both lookbacks. The engine should calculate the minimum warm-up length automatically by introspecting the strategy's indicator declarations, then add a safety buffer of 10-20% to account for any initialization transients.

During the warm-up phase, the backtesting engine must suppress all performance calculations and trade logging. This includes:

• No entries recorded in the trade ledger
• No updates to the equity curve
• No Sharpe Ratio or drawdown calculations
• No benchmark comparison snapshots The engine processes orders and fills silently to advance the state, but the analytics subsystem treats this period as non-existent for reporting purposes.

The warm-up period must be deterministically reproducible for auditability. The engine must record the exact timestamp marking the end of the burn-in phase and the complete state snapshot at that boundary. When replaying a backtest, the system can optionally skip the warm-up phase entirely by loading this serialized state checkpoint, dramatically accelerating iterative strategy development while preserving bit-exact result reproduction.